Methodology and Limitations

Approach

This project uses a descriptive, comparative approach. It does not attempt to estimate causal effects. The goal is to document patterns in the data and to understand what different measures of inequality actually capture.

Where the analysis involves statistical relationships (e.g., Gini vs. life expectancy), these are presented as correlations, not causal claims. The text discusses the causal debate but does not take a strong position.

Key methodological choices

Income concepts

The World Inequality Database provides three income concepts:

Factor income: Labor + capital income before any government intervention
Pre-tax national income: Factor income + pensions and social insurance, before income tax
Post-tax disposable income: After all taxes and cash transfers

The choice of concept matters enormously. Pre-tax inequality is much higher than post-tax inequality in redistributive countries (Nordics) but similar in the US. Most analysis in this project uses both pre-tax and post-tax to show how redistribution changes the picture.

PPP adjustment

Cross-country income comparisons use Purchasing Power Parity (PPP) exchange rates, which adjust for differences in price levels. WID reports income in local currency; we convert using WID’s own PPP factors (xlcuspi999), matched to the year of the income data to avoid inflation distortion.

PPP adjustment is essential for global comparisons but imperfect — it assumes a common consumption basket that may not reflect actual spending patterns in different countries.

Unit of analysis

WID uses equal-split adults as the default unit: household income is divided equally between adult members. This means a single person earning $60k and a couple each earning $30k are treated the same. This is a reasonable default but masks within-household inequality.

Country selection

The core seven countries (US, UK, France, Germany, Sweden, Denmark, Norway) were chosen to represent different positions on the inequality-redistribution spectrum among rich democracies. They are not a random sample of anything.

The expanded set in Notebook 07 adds countries across income tiers but is still a convenience sample, not a comprehensive survey.

Known limitations

WID data quality varies by country. Data for the US, France, and Nordics are based on comprehensive tax records. Data for many developing countries rely on surveys with known undercoverage of top incomes.
Cross-sectional analysis cannot establish causation. The correlation between inequality and poor health outcomes could reflect reverse causation, confounding, or the concavity of the income-health relationship.
The “social wage” is invisible. Nordic countries provide extensive public services (healthcare, education, childcare) that don’t appear in disposable income statistics. Comparing cash income understates the living standards of lower-income Scandinavians relative to Americans.
Happiness data is noisy. The Cantril Ladder is a single survey question with significant measurement error. Cross-country comparisons may be affected by cultural differences in response styles.
Small N problem. With 7 core countries (or 23 in the expanded set), statistical power is limited. A single outlier (the US) can drive or destroy correlations.
Time period sensitivity. Most cross-sectional comparisons use ~2019 data. Results may differ for other years, and the COVID-19 pandemic has disrupted many indicators.

Time-series analysis (Notebook 08)

Notebook 08 examines the Easterlin Paradox: whether countries that get richer over time also get happier. This introduces three additional data sources:

Gallup World Poll / Cantril Ladder (via Our World in Data): ~2005–2024, 130+ countries. The same 0–10 scale used in Notebook 07, but now as a time series rather than a cross-sectional snapshot.
US General Social Survey (GSS): 1972–present. A 3-category happiness question (“very happy,” “pretty happy,” “not too happy”). The canonical dataset for the Easterlin Paradox.
World Values Survey / European Values Study: 1981–2022, ~100 countries surveyed in 7 waves. Life satisfaction on a 1–10 scale. The longest available multi-country time series.

Key limitations specific to the time-series analysis:

Ordinal scales: The GSS 3-point scale cannot distinguish cardinal changes in well-being. Different assumptions about the spacing between categories can produce different trend conclusions (Bond & Lang 2025).
Short vs. long run: Business-cycle fluctuations in GDP are correlated with happiness changes, but sustained long-run growth may not be (Easterlin & O’Connor 2020). Time horizon matters.
WVS coverage is intermittent: Countries are surveyed in waves (~every 5 years), not annually, making precise trend estimation difficult.
Cultural response bias: Cross-country comparisons of subjective scales may be affected by culturally different norms around expressing satisfaction.

Labor market analysis (Notebook 09)

Notebook 09 examines labor force participation and working hours across countries, using three additional data sources:

Our World in Data / Penn World Table (working hours): Average annual hours per person engaged. PWT provides data for 185 countries from 1950 onward; historical data from Huberman & Minns (2007) extends select countries back to 1870.
World Bank (labor force participation): Total, male, and female participation rates from the ILO-modeled estimates. Coverage: 1990-2024, most countries.
Approximate productivity: GDP per hour worked is computed as GDP per capita divided by (labor force participation rate x hours per worker). This is an approximation that assumes employment rate equals participation rate. True GDP-per-hour data from the OECD is more precise but covers fewer countries.

Key limitations specific to the labor analysis:

Economy-wide averages: Average hours per worker masks enormous variation within countries — between full-time and part-time workers, formal and informal sectors, and different occupations.
Historical data covers only production workers: Pre-1950 data from Huberman & Minns measures full-time workers in non-agricultural activities, not the full economy.
Informal work: In developing countries, labor force participation rates undercount informal work, especially women’s unpaid domestic labor. Cross-country comparisons must be interpreted cautiously.
Causation vs. correlation: The relationship between inequality and working hours could run in either direction or be driven by common causes (labor market institutions, tax policy, cultural norms).

Reproducibility

All code is available in the GitHub repository. Data are fetched from public APIs (World Bank, WID, OWID) and cached locally. To reproduce the analysis:

git clone https://github.com/clayborneo/inequality-inquiry.git
cd income-inequality
pip install -r requirements.txt
jupyter nbconvert --to notebook --execute notebooks/*.ipynb

The cache layer (data/cache.py) stores API responses as parquet files to avoid repeated downloads. Clear the cache with python -c "from data.cache import clear_cache; clear_cache()".

--- title: "Methodology and Limitations" --- ## Approach This project uses a **descriptive, comparative** approach. It does not attempt to estimate causal effects. The goal is to document patterns in the data and to understand what different measures of inequality actually capture. Where the analysis involves statistical relationships (e.g., Gini vs. life expectancy), these are presented as correlations, not causal claims. The text discusses the causal debate but does not take a strong position. ## Key methodological choices ### Income concepts The World Inequality Database provides three income concepts: - **Factor income**: Labor + capital income before any government intervention - **Pre-tax national income**: Factor income + pensions and social insurance, before income tax - **Post-tax disposable income**: After all taxes and cash transfers The choice of concept matters enormously. Pre-tax inequality is much higher than post-tax inequality in redistributive countries (Nordics) but similar in the US. Most analysis in this project uses both pre-tax and post-tax to show how redistribution changes the picture. ### PPP adjustment Cross-country income comparisons use Purchasing Power Parity (PPP) exchange rates, which adjust for differences in price levels. WID reports income in local currency; we convert using WID's own PPP factors (`xlcuspi999`), matched to the year of the income data to avoid inflation distortion. PPP adjustment is essential for global comparisons but imperfect — it assumes a common consumption basket that may not reflect actual spending patterns in different countries. ### Unit of analysis WID uses **equal-split adults** as the default unit: household income is divided equally between adult members. This means a single person earning \$60k and a couple each earning \$30k are treated the same. This is a reasonable default but masks within-household inequality. ### Country selection The core seven countries (US, UK, France, Germany, Sweden, Denmark, Norway) were chosen to represent different positions on the inequality-redistribution spectrum among rich democracies. They are not a random sample of anything. The expanded set in Notebook 07 adds countries across income tiers but is still a convenience sample, not a comprehensive survey. ## Known limitations 1. **WID data quality varies by country.** Data for the US, France, and Nordics are based on comprehensive tax records. Data for many developing countries rely on surveys with known undercoverage of top incomes. 2. **Cross-sectional analysis cannot establish causation.** The correlation between inequality and poor health outcomes could reflect reverse causation, confounding, or the concavity of the income-health relationship. 3. **The "social wage" is invisible.** Nordic countries provide extensive public services (healthcare, education, childcare) that don't appear in disposable income statistics. Comparing cash income understates the living standards of lower-income Scandinavians relative to Americans. 4. **Happiness data is noisy.** The Cantril Ladder is a single survey question with significant measurement error. Cross-country comparisons may be affected by cultural differences in response styles. 5. **Small N problem.** With 7 core countries (or 23 in the expanded set), statistical power is limited. A single outlier (the US) can drive or destroy correlations. 6. **Time period sensitivity.** Most cross-sectional comparisons use ~2019 data. Results may differ for other years, and the COVID-19 pandemic has disrupted many indicators. ## Time-series analysis (Notebook 08) Notebook 08 examines the Easterlin Paradox: whether countries that get richer over time also get happier. This introduces three additional data sources: - **Gallup World Poll / Cantril Ladder** (via Our World in Data): ~2005–2024, 130+ countries. The same 0–10 scale used in Notebook 07, but now as a time series rather than a cross-sectional snapshot. - **US General Social Survey (GSS)**: 1972–present. A 3-category happiness question ("very happy," "pretty happy," "not too happy"). The canonical dataset for the Easterlin Paradox. - **World Values Survey / European Values Study**: 1981–2022, ~100 countries surveyed in 7 waves. Life satisfaction on a 1–10 scale. The longest available multi-country time series. Key limitations specific to the time-series analysis: - **Ordinal scales**: The GSS 3-point scale cannot distinguish cardinal changes in well-being. Different assumptions about the spacing between categories can produce different trend conclusions (Bond & Lang 2025). - **Short vs. long run**: Business-cycle fluctuations in GDP are correlated with happiness changes, but sustained long-run growth may not be (Easterlin & O'Connor 2020). Time horizon matters. - **WVS coverage is intermittent**: Countries are surveyed in waves (~every 5 years), not annually, making precise trend estimation difficult. - **Cultural response bias**: Cross-country comparisons of subjective scales may be affected by culturally different norms around expressing satisfaction. ## Labor market analysis (Notebook 09) Notebook 09 examines labor force participation and working hours across countries, using three additional data sources: - **Our World in Data / Penn World Table** (working hours): Average annual hours per person engaged. PWT provides data for 185 countries from 1950 onward; historical data from Huberman & Minns (2007) extends select countries back to 1870. - **World Bank** (labor force participation): Total, male, and female participation rates from the ILO-modeled estimates. Coverage: 1990-2024, most countries. - **Approximate productivity**: GDP per hour worked is computed as GDP per capita divided by (labor force participation rate x hours per worker). This is an approximation that assumes employment rate equals participation rate. True GDP-per-hour data from the OECD is more precise but covers fewer countries. Key limitations specific to the labor analysis: - **Economy-wide averages**: Average hours per worker masks enormous variation within countries — between full-time and part-time workers, formal and informal sectors, and different occupations. - **Historical data covers only production workers**: Pre-1950 data from Huberman & Minns measures full-time workers in non-agricultural activities, not the full economy. - **Informal work**: In developing countries, labor force participation rates undercount informal work, especially women's unpaid domestic labor. Cross-country comparisons must be interpreted cautiously. - **Causation vs. correlation**: The relationship between inequality and working hours could run in either direction or be driven by common causes (labor market institutions, tax policy, cultural norms). ## Reproducibility All code is available in the [GitHub repository](https://github.com/clayborneo/inequality-inquiry). Data are fetched from public APIs (World Bank, WID, OWID) and cached locally. To reproduce the analysis: ```bash git clone https://github.com/clayborneo/inequality-inquiry.git cd income-inequality pip install -r requirements.txt jupyter nbconvert --to notebook --execute notebooks/*.ipynb ``` The cache layer (`data/cache.py`) stores API responses as parquet files to avoid repeated downloads. Clear the cache with `python -c "from data.cache import clear_cache; clear_cache()"`.